Multi-condition building decarbonization using deep reinforcement learning and large language model

Abstract

The subjectivity of building management and the lack of human–machine interaction make building operation decarbonization challenging. This work designs a building control-oriented optimization framework to reduce carbon emissions and automatically produce the strategies. Firstly, building information modeling is constructed by referring to on-site data and weather conditions to create a multi-condition dataset. Secondly, the surrogate models regarding carbon emissions and thermal comforts correspond to 22-30℃, 30-35℃, and above 35℃ during the hot season. Thirdly, the optimal decarbonization strategy under different weather conditions is identified by deep reinforcement learning. Finally, a human–machine interactive interface is developed to display the results and provide valuable suggestions. The proposed decarbonization framework has been validated in a green building in China, and the results reveal that: (1) The building information model can accurately simulate actual carbon emissions with an error of 0.1%. (2) The surrogate models show excellent prediction for carbon emissions and thermal comforts with an R² of 0.93 in the testing sets. (3) The optimization rates corresponding to 22-30℃, 30-35℃, and above 35℃ are 47.28%, 17.75%, and 13.58%, respectively, and the decarbonization-based LLM can provide practical strategy according to outdoor temperature and user preferences. The work contributes to weather-based building control optimization and the development of a building decarbonization large language interaction model.